Comparing Four Frailty Index Construction Methods for Mortality Prediction In the Health and Retirement Study

Background and Objectives: The frailty index of cumulative deficits (FI-CD) is widely used to quantify frailty as the proportion of observed health deficits, yet its equal-weighting assumption has been criticized for treating all deficits as interchangeable. Principal component analysis (PCA) and item response theory (IRT) have emerged as leading alternatives for constructing empirically weighted indices. However, existing comparisons have focused primarily on predictive performance, with limited attention to score concordance across demographic subgroups or consistency in identifying the most informative deficits. This study systematically compared four frailty indices—two cumulative-deficit versions (FI-CD60, FI-CD35), FI-PCA, and FI-IRT—on inter-method agreement, subgroup invariance, item-level consistency, and predictive validity for all-cause mortality.

Methods: Using data from the Health and Retirement Study (1998–2022), we conducted a longitudinal analysis of adults aged 50 and older. 60 health deficit variables were retained after excluding items with more than 5% missing data and those too rare (<1%) or too common (>80%) to differentiate frailty levels. To avoid redundancy among conceptually overlapping items, those sharing similar content were combined into ordinal scales, yielding 35 analytic items. Missing covariates were addressed via listwise deletion; missing deficit values were handled through multiple imputation by chained equations (m = 5), with FI-CD35, FI-PCA, and FI-IRT scores averaged across imputed datasets. The final sample comprised 192,578 person-wave observations from 35,403 individuals. FI-CD60 and FI-CD35 were calculated as mean deficit proportions, FI-PCA was derived from polychoric PCA retaining 5 components identified by parallel analysis, and FI-IRT was estimated using graded response models with scores rescaled to 0–1. Inter-method concordance was assessed using Pearson and Spearman correlations, computed overall and by sex, age, and ethnicity. Predictive validity for all-cause mortality was evaluated with survey-weighted discrete-time complementary log–log models, adjusting for sex, age, ethnicity, education, smoking, and alcohol use.

Results: FI-CD60 (mean = 0.168, SD = 0.128), FI-CD35 (mean = 0.203, SD = 0.149), and FI-PCA (mean = 0.179, SD = 0.153) displayed positively skewed distributions, whereas FI-IRT (mean = 0.477, SD = 0.151) approximated a bell-shaped distribution. All four indices showed high concordance: Pearson correlations ranged from 0.921 to 0.985 and Spearman correlations from 0.958 to 0.990, with comparable magnitudes across subgroups (Pearson: 0.915–0.987; Spearman: 0.949–0.993). PCA and IRT exhibited strong item-level consistency, identifying the same top 4 most informative deficits: bathing difficulty, housework limitation, impairment-related limitation, and walking difficulty. Adjusted mortality models yielded near-equivalent predictive accuracy: pseudo-R² values were 0.217 for FI-CD60, 0.215 for FI-CD35, 0.216 for FI-PCA, and 0.218 for FI-IRT.

Conclusions and Implications: The four frailty indices produced near-equivalent mortality predictive accuracy and high inter-method concordance across demographic subgroups, suggesting that cumulative-deficit scores remain strong benchmarks for aging research. Although FI-CD performs well for risk stratification and service planning, researchers seeking to understand which deficits contribute most to frailty severity may benefit from applying PCA or IRT, as both methods consistently identified functional dependency items as carrying the greatest weight. These findings advance methodological rigor in health measurement and clarify when added modeling complexity is warranted.